Google DeepMind Releases A Handy Tool To Detect Text Written By AI

Google DeepMind announced it has developed a tool for identifying AI-generated text, and is making it available via open source. SynthID is a technology from Google DeepMind that watermarks and identifies AI-generated content by embedding digital watermarks directly into AI-generated images, audio, text, or video.

A lot of AI generated content, such as Toys R Us’ AI-generated origin story video trailer, have become very good, making it harder to determine whether humans or AI made the content. Because of this, many companies and people are looking for easy ways of being able to tell the difference between AI-generated content, and that created by humans. The DeepMind team has been working on providing effective and easy methods of being able to detect AI-generated content, announcing watermarks for images last year, and later for AI-generated video. Now, the company has unveiled another tool for helping identify AI-generated content, this time for text.

“Now, other [generative] AI developers will be able to use this technology to help them detect whether text outputs have come from their own [large language models], making it easier for more developers to build AI responsibly,” remarked Pushmeet Kohli, the vice president of research at Google DeepMind. 

Watermarks are a recognizable image or pattern embedded into a digital or physical document to identify the owner, prevent unauthorized duplication, or add a decorative effect. They have been around for a while, with many photographers using them to safeguard their work. SynthID works by adding an invisible watermark directly into AI generated text.

In a press release, Google remarked, “Practically speaking, SynthID Text is a logits processor, applied to a model’s generation pipeline after Top-K and Top-P, that augments the model’s logits using a pseudorandom g-function to encode watermarking information in a way that balances generation quality with watermark detectability.”

Kohli explained the technology introduces additional information at the point of generation by changing the probability that tokens will be generated. In order to detect the watermark and determine whether the text is AI-generated, SynthID compares the expected probability scores for words in watermarked and unwatermarked text.

While Google says SynthID does not compromise the quality, accuracy, creativity, or speed of generated text, it does have limitations. The watermark proved effective to some forms of tampering, such as cropping text, and light editing or rewriting. However, it was less reliable when it came to AI-generated text that had been written or translated from one language to another. It was also found to be less reliable in responses to prompts asking for factual information, such as asking for the capital of a country.

“Achieving reliable and imperceptible watermarking of AI-generated text is fundamentally challenging, especially in scenarios where LLM outputs are near deterministic, such as factual questions or code generation tasks,” remarked Soheil Feizi, an associate professor at the University of Maryland.  

Because of all the possible downsides to technologies such as DeepMind’s SynthID, many are saying watermarks can only be one aspect of the solution. Irene Solaiman, Hugging Face’s head of global policy, remarked, “Watermarking is one aspect of safer models in an ecosystem that needs many complementing safeguards. As a parallel, even for human-generated content, fact-checking has varying effectiveness.”